The two trust problems of AI

Royden James, Chief Product &Technology Officer, Viedoc

We call too many things AI

If there's one thing I've learned over the past few years, it's that we use the term "AI" to describe far too many different things.

A chatbot is called AI. A predictive model is called AI. Workflow automation is called AI. Natural language search is called AI. Sometimes even software that's simply applying well-defined rules gets described that way.

The result is that conversations about AI often become conversations about a single, imaginary technology that somehow understands language, reasons like a human, automates complex workflows, and makes decisions. That's not the reality.

I don't come from clinical research originally. I come from building software products. Looking at this industry with fresh eyes, what strikes me isn't whether AI will transform clinical trials. I think it will. It's that we often struggle to separate what AI is actually doing from what well-engineered software has always done.

That distinction matters because it changes where trust belongs.

AI is the interface, not the system

Today, most people experience AI through conversation. You describe what you're trying to accomplish instead of navigating menus, learning complex systems, or writing technical queries. That's a meaningful shift because understanding language is something modern AI does exceptionally well.

But understanding language isn't the same as doing the work.

When a system retrieves data, validates information, applies protocol rules, generates a report, or completes a workflow, those actions depend on software that has been deliberately designed, tested, and validated. The AI is making that capability easier to access by translating human intent into something the system understands.

In many ways, AI is becoming an interface rather than the system itself.

That distinction changes the questions we should ask.

The two trust questions

Instead of asking whether we trust AI, we're really asking two different questions.

First, can the AI correctly understand what I'm trying to do?

Second, can I trust the system carrying out the work?

Those are different problems with different answers.

The first depends largely on advances in language models. The second depends on product design, engineering, validation, governance, and compliance. That's where software companies earn trust.

It's also where responsibility lies.

Where automation should (and shouldn't) go

Too often, discussions about AI jump straight to whether machines will replace experts. I think that's the wrong conversation.

The real opportunity is to reduce unnecessary work, not expertise.

There's a difference between removing friction and removing judgment.

Many repetitive tasks can be automated safely because they're deterministic, well understood, and appropriately validated. Other activities require interpretation, experience, or accountability and should continue to involve human oversight. The important question isn't whether a human is involved. It's whether the level of automation matches the level of risk.

That's the principle that should guide how AI is introduced into regulated industries.

Ironically, the better these systems become, the less people should have to think about the AI itself.

Ask this instead

The conversation should shift from asking, "Do you have AI?" to questions like:

• What work disappears because of it?
• Which parts of this workflow are automated, and why?
• What validation supports those automations?
• How are actions audited?
• Where is human oversight intentionally retained?

Those questions reveal far more than whether a platform includes a chatbot.

They also make it easier to separate trust in AI from trust in the software around it.

The real question

AI should be trusted to understand language within its capabilities. Software should be trusted because it's engineered, validated, transparent, and appropriate for the task it's performing. Those aren't competing ideas. They're complementary responsibilities.

Clinical research has always been built on evidence, traceability, and proportional risk management. AI shouldn't change those principles. It should make them easier to work with.

The most valuable AI won't be the most visible. It will quietly reduce the effort required to move from a question to an answer while preserving the controls, validation, and oversight that regulated research demands.

The future of AI in clinical research won't be defined by how intelligent the interface appears. It will be defined by how trustworthy the systems behind that interface are.

That's why I think we're asking the wrong question.

The question isn't whether a platform has AI.

The question is: what part of the solution am I being asked to trust?